# roadkeep-gui changelog — alegauss

Every entry in roadkeep-gui's roadkeep ledger, dated by the commit that wrote it.

- 2026-09-24 · RG304: Enter in a session's reply box never sends it, so every answer takes a trip to the button
- 2026-09-23 · RG307: the app has no public page, so a release is reachable only by somebody who already knows the repository
- 2026-09-21 · RG303: a reply typed into a session is never drawn in its stream, so what was asked of it is gone once sent
- 2026-09-21 · RG302: show refuses every shipped entry: its reader holds the pointer to a string, and a ledger line carries none
- 2026-09-21 · RG301: a walkthrough is asked through askGloss, so the module that runs every read-only question is named after one of them
- 2026-09-21 · RG300: no shot fixture keeps an answer, so a kept gloss's notices and the walkthrough dialog are drawn by no picture
- 2026-09-21 · RG299: an explanation nests a paragraph inside a paragraph, which React reports as invalid HTML on every answer
- 2026-09-21 · RG298: npm run dev:inspect opens its first window without the debugging port, so Playwright MCP reaches it only after a restart
- 2026-09-21 · RG297: an explanation being written shows only the last files read, so a slow run cannot be told from a stuck one
- 2026-09-21 · RG296: a project row says what is open and nothing about what shipped with nobody looking at it
- 2026-09-21 · RG295: a validation that failed ends in a sentence nobody files, so the defect it found is lost at the dialog
- 2026-09-21 · RG294: a person who followed a walkthrough has nowhere to say what happened, so the verdict never reaches the ledger
- 2026-09-21 · RG293: nothing on a project surface says what is shipped and unlooked-at, so the list lives in a terminal
- 2026-09-21 · RG292: a walkthrough is lost when the sheet closes, and the sheet is closed by the person going off to follow it
- 2026-09-21 · RG291: a person told a line is unvalidated has to work out how to check it from one sentence in the ledger
- 2026-09-21 · RG290: a gloss kept before an answer grew a slot says nothing under it, and reads as current
- 2026-09-21 · RG289: the app can neither ask which entries await a person nor record a verdict: neither verb is in its tables
- 2026-09-21 · RG288: an explanation answers from the brief alone, so it cannot say where in the code the work lands or what it will touch
- 2026-09-21 · RG287: an explanation is lost when its dialog closes, so reopening a task pays the whole wait and the tokens again
- 2026-09-21 · RG286: an explanation reads as a column of paragraphs, while a task's chain, its before and after and its steps are shapes
- 2026-09-21 · RG285: the task screen offers no plain account of a task, only the line and a design written for whoever builds it
- 2026-09-21 · RG284: the only door that puts a task before Claude Code is a session that may write, too much to ask what a task means
- 2026-09-21 · RG283: nothing asks Claude Code what a task means in plain words, so a line and its design are read in their own terse terms
- 2026-09-21 · RG282: an edited file opens as it is now, with no view of it against what it was before the session
- 2026-09-21 · RG281: a file that moved on disk while the session ran never says whether it appeared, changed or went away
- 2026-09-21 · RG280: an edited file's row never says whether the session created, changed or deleted it
- 2026-09-21 · RG279: the session screen's side bars are 18rem whatever they hold, so a long path is cut and the stream cannot be widened
- 2026-09-21 · RG277: the session screen's side cards stay where the grid puts them, and cannot be dragged to the other side or reordered
- 2026-09-21 · RG276: the settings file has nowhere to keep how the session screen's cards are arranged, so none outlives the window
- 2026-09-21 · RG275: a session a task hands over grows the page with every act, where one the gate hands over scrolls in its stream
- 2026-09-17 · RG273: a session runs on a hand-written copy of Claude Code's stream protocol, so each change to it is this app's to chase
- 2026-09-17 · RG272: a running session is refused every call that needs permission, and the person watching is never asked
- 2026-09-17 · RG271: a session's words are drawn with their Markdown literal, so bold, code and lists read as asterisks and backticks
- 2026-09-17 · RG270: a tool call the session was refused is never shown, and nothing lets a person allow it for the next turn
- 2026-09-17 · RG269: a session that stopped to ask something cannot be answered from the window, only from a terminal with its id
- 2026-09-17 · RG268: a session that ends its turn waiting on a person is drawn as done, so a stalled task reads as finished
- 2026-09-17 · RG267: the shortcuts sheet lists / for the palette, which nothing binds, and a palette opened by its key searches nothing
- 2026-09-17 · RG266: leading from an edit to its act is never checked to scroll, and throws unhandled in every unit run
- 2026-09-17 · RG265: the files a session touched are two flat lists of full paths, sharing a panel with what moved in the backlog
- 2026-09-17 · RG264: a session handed a gate finding is drawn in the reading column, since the width table names one session route
- 2026-09-16 · RG263: a door that reads prose has only a box, so the reader who cannot write it has no way to hand it over
- 2026-09-16 · RG262: the running gate is a sentence that never changes, so a slow run and a window that stopped read the same
- 2026-09-16 · RG261: a door whose value is - reads prose on standard input, which no screen asks for and no call ever closes
- 2026-09-15 · RG260: the window's gate run and a door's carry no deadline, so a call that hangs leaves the screen running for good
- 2026-09-15 · RG259: a door's button says Take it, so the control that runs a command line reads as accepting the finding
- 2026-09-15 · RG258: a finding's code is a bare token, so what a class like ref.unresolved means is still looked up in a terminal
- 2026-09-15 · RG257: the Portuguese verdict divergente names no check, so a reader takes six findings for six items that diverged
- 2026-09-15 · RG256: a row's gate verdict is plain text, so the findings it counts are reached only by guessing Run the gate holds them
- 2026-09-15 · RG255: the project screen says nothing of its gate, so findings counted on its portfolio row vanish one click later
- 2026-09-15 · RG254: the gate opens on a count when its verdict is fresh, so the findings it counts appear only after a second run
- 2026-09-15 · RG253: gate verdicts are forgotten at quit, so every launch runs lint once per project before a row has a verdict
- 2026-09-15 · RG252: a remembered row is still read in full behind the screen, so a launch spends every interpreter a cold start did
- 2026-09-15 · RG251: every launch reads every backlog again, so the portfolio shows skeletons until the slowest engine answers
- 2026-09-15 · RG250: a launch starts every project's engine at the same moment, so twenty projects pin the machine together
- 2026-09-15 · RG249: a portfolio read has no deadline, so one project whose engine hangs keeps every row's next line pending
- 2026-09-15 · RG248: a walk that lands after the list is drawn sets every row back to pending, so the portfolio loads twice
- 2026-09-15 · RG247: files a session changed through a command and not an edit call appear nowhere on the session screen
- 2026-09-15 · RG246: the file viewer shows a file as it is now and not what the session changed in it
- 2026-09-15 · RG245: a file the session edited cannot be opened from the window, so reading it means leaving for an editor
- 2026-09-15 · RG244: an edited file is listed on the session's word, so a call that reported success and changed nothing looks the same
- 2026-09-15 · RG243: the session screen does not list the files the agent edited, so what it changed is found by reading the stream
- 2026-09-15 · RG242: the task and session screens do not show which project they belong to, so finding out means going back
- 2026-09-15 · RG241: the order a person picks for the portfolio is gone at the next launch, so the table opens in the record's order again
- 2026-09-15 · RG240: the portfolio cannot rank projects by how many lines they have open, so the heaviest backlog is found by reading
- 2026-09-15 · RG239: the portfolio lists projects only in the record's order, so finding one by name means reading every row
- 2026-09-15 · RG238: the ground control in the header is a sentence, where Turing wears the design system's icon toggle
- 2026-09-15 · RG237: the session is held to the reading column, so its stream between two side columns is left about 360 pixels
- 2026-09-15 · RG236: npm run dev and npm start open a window wearing Electron's atom, though build/icon.svg is the app's mark
- 2026-09-15 · RG235: on a task, Copy the brief sits higher than Hand to Claude Code beside it, so the hero's two actions share no line
- 2026-09-14 · RG234: the fast suite re-transforms every module on every run, and nothing caches that between runs
- 2026-09-14 · RG233: CLAUDE.md and the roadmap skill give opposite commit rules, and both are loaded every turn
- 2026-09-14 · RG232: a task test failed once inside a full run and left no message, so nobody can tell what it waited on
- 2026-09-14 · RG231: the shortcuts button still shows at 400 wide, which RG215 meant to hide
- 2026-09-14 · RG230: a responsive class beside a base class the design system ships is overridden at every width, and no gate sees the pair
- 2026-09-14 · RG229: at 1280 a project's lines are stacked into one column, which RG223 meant only for phone width
- 2026-09-14 · RG228: at 400 wide the sessions list draws its headings over each other and cuts each project name to a letter
- 2026-09-14 · RG227: the shell drawing's rail draws five tiles against the window's three, and no rail tile is marked
- 2026-09-14 · RG226: tests fail inside a full run and pass on their own, so a red gate can mean the machine was busy
- 2026-09-14 · RG225: at 400 wide the session opens on two panels, so the stream a reader came for starts below the fold
- 2026-09-14 · RG224: the phone-width run draws every surface in English, so a sentence that overflows only in Portuguese passes
- 2026-09-14 · RG223: a line's symptom is one letter per line at 400 wide, the column left about ten pixels
- 2026-09-13 · RG222: npm run shots builds the shell alone, so a changed screen is photographed from an older renderer bundle
- 2026-09-13 · RG221: the rail draws only Home, so the settings and the sessions are reached through the palette alone
- 2026-09-13 · RG217: the stream region is taller than the space under the hero, so its end and the jump control sit below the fold
- 2026-09-13 · RG216: a count of one is written as a plural, so the portfolio says 1 projetos and 1 lidos
- 2026-09-13 · RG215: the header and a hero's actions do not fit at 400 wide, so every surface scrolls sideways at phone width
- 2026-09-13 · RG214: no test asks whether a surface scrolls sideways at 400 wide, or whether Tab walks its controls in reading order
- 2026-09-12 · RG220: a screenshot launch that fails before the window draws leaves its Electron process running
- 2026-09-12 · RG219: the detail and graph tests hunt this backlog for a line waiting on outside work, so a backlog without one reddens them
- 2026-09-12 · RG218: the release tag test runs whichever bash is first on PATH, and on Windows that can be WSL's, which has no node
- 2026-09-12 · RG213: the renderer's tests run in jsdom alone, so a test about scroll, overflow or focus order asserts numbers it invented
- 2026-09-12 · RG212: an agent checking a screen improvises its own harness, and nothing says what to look for once it has the pictures
- 2026-09-12 · RG211: no accessibility check runs over a rendered surface, so a defect only real layout and colour show ships green
- 2026-09-12 · RG210: the session screen only draws behind a real Claude Code run, so its stream, scroll and folded notes are checked blind
- 2026-09-12 · RG209: no command shows what a screen looks like, so an agent checks a screen change with tests that never see the window
- 2026-09-12 · RG208: the session stream draws every system note, init to token counts, between the acts a reader follows
- 2026-09-12 · RG207: no settings screen, so the ground and the language are loose header controls and a new preference has nowhere to go
- 2026-09-12 · RG206: the session stream does not follow new acts, so a running task is read by scrolling the page by hand
- 2026-09-12 · RG205: a session bills an ANTHROPIC_API_KEY the machine carries, not its Claude Code login, and dies for lack of credit
- 2026-09-12 · RG204: an emoji is the only icon a row can draw, so a project with a logo file cannot show it
- 2026-09-12 · RG203: a declared name is read from the checkout, so a project reverts to its folder name as it goes missing
- 2026-09-12 · RG202: five screens still name a project by its folder, so a declared name vanishes the moment one is opened
- 2026-09-12 · RG201: a row says a path and never what the project is for, so seventeen rows read as seventeen paths
- 2026-09-12 · RG200: every project chip is the same folder glyph tinted by a hash, so no row is recognisable at a glance
- 2026-09-12 · RG199: nothing on a row says which branch a worktree is on, so a declared name renders a family as identical rows
- 2026-09-12 · RG198: a row is named by its folder, so two worktrees of one product read as 2026.3 and 2026.2
- 2026-09-12 · RG197: a live row test needs this backlog to be a different size from a fixture, and shipping made them equal
- 2026-09-12 · RG196: the graph test hunts this backlog for a blocked line, so draining the backlog reddens three assertions
- 2026-09-12 · RG195: the fixture hands out an unresolved temp path, so a live assertion fails on a Windows account with a short name
- 2026-09-12 · RG194: no public page says who may request a signature and what gets signed, so the certificate cannot be applied for
- 2026-09-12 · RG193: an engine too old for `config` reads as an unreadable answer, so a row shows a usage dump and no version
- 2026-09-12 · RG190: a handover landing after the reader left the task screen moves the window to the session anyway
- 2026-09-12 · RG170: list prints no readiness per line, so every row costs a deps read and startable only cannot be offered
- 2026-09-12 · RG162: no LICENSE file ships and every build says 0.0.0, so the first release the maintainer decided on cannot be cut
- 2026-09-12 · RG154: the update check has only met answers written by hand, never GitHub's own for a published release
- 2026-09-12 · RG147: the palette says it finds a line in every backlog and lists only surfaces, so search over all projects reaches no one
- 2026-09-12 · RG128: a checkout busy when the probe runs is skipped, so the launcher answers from a cache 446 versions old
- 2026-09-11 · RG192: two screens draw the English a failed bridge call carried, a field whose code no catalogue entry covers
- 2026-09-11 · RG191: a screen drawing a field this app wrote in English passes every run that does not reach the state showing it
- 2026-09-11 · RG189: a cache entry refreshed against a new stamp keeps its old place, so eviction takes the entries being used most
- 2026-09-11 · RG188: a payload key spelled __proto__ is dropped by the dictionary reader, so a count vanishes and the read still answers ok
- 2026-09-11 · RG187: the carrier gates every project as it opens, so a cold start over seventeen starts seventeen lints at once
- 2026-09-11 · RG186: the duplicates gate matches on export, so a screen redeclaring a design-system name for its own use passes it
- 2026-09-11 · RG185: the gate surface runs lint every time it opens, though the ledger already holds a verdict the files have not moved under
- 2026-09-11 · RG184: a source file carrying control characters passes every gate, so an invisible byte reaches a commit
- 2026-09-11 · RG183: a catalogue key no screen says is never reported, so a string both locales declare can be dead
- 2026-09-11 · RG182: the filing form asks nobody what the block already delivered, so the window spends an id the terminal would question
- 2026-09-11 · RG181: the door keep holds every batch a session offered, since only a file moving drops one
- 2026-09-11 · RG180: nothing says when the walk behind the remembered record landed, so a project found or lost shows on the next ask
- 2026-09-11 · RG179: an amend that changes deps or requires reads as unreadable, though the file was already written
- 2026-09-11 · RG178: the sessions list asks once when it opens, so one that ends while it stands still reads as running
- 2026-09-11 · RG177: a time is written in the desktop's locale and not the window's, so a Portuguese window dates a file in English
- 2026-09-11 · RG176: the pseudo-locale run draws the portfolio alone, so three of this window's four surfaces are never read for a literal
- 2026-09-11 · RG175: a line this window handed over is offered to a second session, since the claim it took names nobody
- 2026-09-11 · RG174: a line's own criteria never reach its task screen, so what the ship must check reads as the block's finish line alone
- 2026-09-11 · RG173: a dep or a chain hop on the task screen is text, so reaching the line it names means going back through the project
- 2026-09-11 · RG172: a listing narrower than its file, and a failed update check, are explained in English whatever the window speaks
- 2026-09-11 · RG171: the improvements tab names each design by its pointer and not its heading, so two designs look alike until opened
- 2026-09-11 · RG169: a root is added two levels deep and no control changes it, so worktrees three levels down still need the settings file
- 2026-09-11 · RG168: an unreadable row's reason is composed in English in core, so a Portuguese window shows it untranslated
- 2026-09-11 · RG167: a row is read once, so a line shipped in a terminal leaves the portfolio stale until the window reopens
- 2026-09-11 · RG166: every row's gate reads unknown, since the ledger RG18 built is fed by nothing a window runs
- 2026-09-11 · RG165: a door the engine offers cannot run from a window, since the bridge refuses a verb or option its tables never compose
- 2026-09-11 · RG164: the carrier keeps its catalogue in memory, so RG14's record is never written and every launch walks every root first
- 2026-09-11 · RG163: three live tests need this backlog to hold a quoted lead, an idea and a prose-only design, and it holds none
- 2026-09-11 · RG161: the signing procedure breaks the Windows CI job on one route and leaves the app exe unsigned on another
- 2026-09-11 · RG160: openHere's own unheld wiring is never run by a test, since RG137's live case rebuilds it by hand
- 2026-09-11 · RG159: the empty-brief contract asserts lacking on a fixture that never has any, so no lacking line is ever read
- 2026-09-11 · RG158: two live assertions pass unchecked: a picked line with no tier, and a combined filter that selects nothing
- 2026-09-11 · RG157: the tag-against-manifest check is tested by its text, so an inverted condition or a lost exit still passes
- 2026-09-11 · RG156: the update check tells a build ahead of every release it is the newest, and hides why a check failed
- 2026-09-11 · RG155: the release-page guard compares a string prefix, so a release URL with dot segments opens another repository
- 2026-09-11 · RG153: nothing in the window hands a task to Claude Code or shows the session, so block F's acts and landings reach no one
- 2026-09-11 · RG152: Run the gate is a button with nothing behind it, so a finding and the doors it names are still read in a terminal
- 2026-09-11 · RG151: no screen composes a write, so a line is still filed in a terminal though its budget, argv and doors are built
- 2026-09-11 · RG150: a task opens nowhere in the window, so brief's join of deps, design and binding lists is read by nobody
- 2026-09-11 · RG149: the changelog, decisions, improvements and deferred tabs are drawn as labels with nothing behind them
- 2026-09-11 · RG148: a project cannot be opened in the window, so its lines, blocks and filters exist only as readers in core
- 2026-09-11 · RG146: roots are read and never written, so Add a root and Rescan roots are buttons nothing can honour
- 2026-09-11 · RG145: the window opens on a scaffold naming three packages, so the portfolio core already computes is drawn nowhere
- 2026-09-11 · RG144: the bridge only answers when asked, so a session's stream and a governed file changing have no way to reach a screen
- 2026-09-11 · RG143: the bridge carries no roadkeep call, so no screen can ask a project anything and every reader in core runs only in tests
- 2026-09-11 · RG77: a non-goal arrives as a lead alone, so the sentence that argues it stays in the file
- 2026-09-11 · RG28: the --stale payload carries each pause's age and reason and the reader drops both, so the store draws in file order
- 2026-09-10 · RG142: a brief with no id on a backlog with nothing ready answers empty, and the reader calls that this app being behind
- 2026-09-10 · RG141: a backlog with nothing ready answers pick with a null tier, so the portfolio draws a waiting project as unreadable
- 2026-09-10 · RG140: the pseudo-locale run reads the text on screen and never a name, so an aria-label typed into a component passes it
- 2026-09-10 · RG139: the executable names GitHub, Inc. as its company, because the manifest names no author to replace Electron's
- 2026-09-10 · RG138: the packaged app wears Electron's icon in the taskbar and installer, though roadkeep's mark now leads its header
- 2026-09-10 · RG137: an open project does not carry why its engine could not be held, so a row cannot say why it reads slowly
- 2026-09-10 · RG136: the artboards draw a brand mark beside the wordmark that the window does not render, and nothing says which is right
- 2026-09-10 · RG135: a project whose engine cannot hold a session reads at spawn speed, and nothing says the held surface was refused
- 2026-09-10 · RG134: the contract reads this repository's own lint for a finding, so a clean gate leaves it nothing to assert
- 2026-09-10 · RG133: the renderer ships three icon API endpoints, and one prop on a nav item is what starts fetching from them
- 2026-09-10 · RG132: a dialog says Close in English in every language, and the run that found it can only name it
- 2026-09-10 · RG131: a read cancelled after it reaches the held engine runs to the end, and only the transport that spawns kills one
- 2026-09-10 · RG130: the live gate runs one file at a time because every read in it spawns a python, and that is most of ten minutes
- 2026-09-10 · RG129: the verb table is declared with string keys, so the file owning it asserts its own key type and nothing checks the pair
- 2026-09-10 · RG127: the artboards draw a header the window no longer has, and they are what the next screen is designed against
- 2026-09-10 · RG126: the provider stack is spelled twice, so one added to the window is missing from every test that renders it
- 2026-09-10 · RG125: the i18next half of the wording is two hand-written objects, so a key added to one and missed in the other is silent
- 2026-09-10 · RG124: a design system control draws its flags from a CDN, so a screen adopting one shows the image the policy refused
- 2026-09-10 · RG123: a sentence shown in a toast is English whatever the window is speaking, and the run that would catch it looks elsewhere
- 2026-09-10 · RG122: the transport that reads in 6ms is proven by a test and reached by nothing a person would run
- 2026-09-10 · RG121: the one type-aware rule left off covers the whole tree, so an unsafe assertion outside a reader is unreported too
- 2026-09-10 · RG90: muted text on a muted panel is 4.33 to 1, which no screen renders yet and one will
- 2026-09-10 · RG50: nothing updates, so a fix reaches a person only if they go and look for one
- 2026-09-10 · RG13: a candidate is judged governed by a read that fails, so rejecting a folder costs the work of a whole lint
- 2026-09-10 · RG9: prose a person typed goes out through argv, which is where RK1474 recorded bytes arriving as different bytes
- 2026-09-09 · RG120: the wired skill and hook are behind the engine answering here, and only a command somebody remembers fixes it
- 2026-09-09 · RG118: no screen names the build, so a defect report cannot say which one it came from
- 2026-09-09 · RG117: the route table is spelled three times, so a surface can be routed and still be missing from the map
- 2026-09-09 · RG116: there is no control for the language, so choosing one means editing the settings file by hand
- 2026-09-09 · RG115: the sentences a bad settings file composes cross the bridge and no screen shows one
- 2026-09-09 · RG114: the built app never starts on the Linux runner, and the harness reports the wait without what it printed
- 2026-09-09 · RG113: the live tests find an engine on the developer's PATH, so a fixture is unresolved on a runner
- 2026-09-09 · RG112: nothing pins the line endings, so a Windows runner checks out CRLF and Prettier reports every text file
- 2026-09-09 · RG111: the lint gate reads the engine CI checks out beside the packages and fails on code this repo did not write
- 2026-09-09 · RG110: the vendored page reference is checked by nothing, so a copy that drifts is read with the same trust as a current one
- 2026-09-09 · RG109: the five screens drawn here miss four things the shell's set requires, one of which stopped being true at RG86
- 2026-09-09 · RG108: the design system's page contract reaches this repository only if somebody goes and reads it in another checkout
- 2026-09-09 · RG107: the dark focus ring is the light accent, because this app's :root wins over the package's dark block
- 2026-09-09 · RG106: the window mounts nothing until the bridge answers, so a call that never returns leaves it blank
- 2026-09-09 · RG105: the rail marks where you are in the package's neutral, so the shell reads as another product's console
- 2026-09-09 · RG104: removing a fixture directory throws EPERM on Windows, so a file whose tests all passed reports as a failed suite
- 2026-09-09 · RG102: the walk yields now and the three reads around it do not, over the same paths a person named
- 2026-09-09 · RG101: every call spends a second and a half starting Python, and the engine publishes a mode that starts once
- 2026-09-09 · RG100: an id in a project whose listing was bounded reads as one nothing in that project ever carried
- 2026-09-09 · RG99: a read has two doors and the one a screen has to use is the one that raises rather than answering
- 2026-09-09 · RG98: core's boundary test holds an import and nothing else, so a path rule written as string work passes it
- 2026-09-09 · RG97: nothing declares which Node this repo needs, so a wrong one fails inside a dependency and not at the door
- 2026-09-09 · RG96: the suite now tests a bundle it did not build, so a green run can be about yesterday's code
- 2026-09-09 · RG95: no gate reads an advisory, so the next one is found by whoever happens to install something
- 2026-09-09 · RG94: no type-aware rule runs, so a floating promise is still nobody's finding
- 2026-09-09 · RG93: the dev loop reaches into the compiler's own layout, and only a hand run would notice it moving
- 2026-09-09 · RG92: the project skill states counts and a priority that stopped being true, and nothing checks it
- 2026-09-09 · RG91: the packaged build is proven on one machine, and CI never makes one
- 2026-09-09 · RG89: two readings derive the marker set from one config, and only a test keeps them agreeing
- 2026-09-09 · RG88: one screen has two translation systems, and only one of them was chosen
- 2026-09-09 · RG87: the ground setting has two homes, and the file one is not the one that wins
- 2026-09-09 · RG86: the locale setting reaches no screen, so choosing one changes nothing
- 2026-09-09 · RG85: the handler proving the seam runs any argv against any root, which a real service cannot
- 2026-09-09 · RG84: every live file resolves the engine itself, so one being rebuilt mid-run reds the suite four ways
- 2026-09-09 · RG83: one verb answers in two shapes under a flag, and nothing names the second until it breaks
- 2026-09-09 · RG82: the contract test asserts one engine version across three reads, so a rebuilt engine reds a suite nothing touched
- 2026-09-09 · RG81: the read and write paths each spell the same argv wrapper, and the two have already drifted apart
- 2026-09-09 · RG80: opening a paused task is a refusal whose typed fields are all empty, leaving the screen only English
- 2026-09-09 · RG78: every live test file now carries its own call-and-unwrap, and the sentence a moved key prints is written in each
- 2026-09-09 · RG76: brief already sends the chains and the shape drops them, so opening one task pays for a second call to deps
- 2026-09-09 · RG75: two live tests still name RG9, so the commit that ships it fails a suite nothing changed to break
- 2026-09-09 · RG63: there is no page chrome, so the first real screen will invent a shape the design system already has
- 2026-09-08 · RG103: every caller assembles the engine, the pool, the cache and the reader itself, and no two assemble it alike
- 2026-09-08 · RG74: a marker says in progress and the claim registry says nobody holds it, and nothing tells the two apart
- 2026-09-08 · RG73: a row is built by four separate reads, so a project costs four interpreter starts to show once
- 2026-09-08 · RG72: a family is drawn in the order it was found, so a version list can read newest last or newest first by luck
- 2026-09-08 · RG71: a walk of every root blocks the thread it runs on, so a slow drive freezes the window
- 2026-09-08 · RG68: the contract costs a minute of process starts, because every read rebuilds nothing but pays a fresh engine
- 2026-09-08 · RG67: a project declaring a read bound answers in a shape nothing here has ever seen, and the reader refuses it
- 2026-09-08 · RG66: a verb can be called with no shape declared for its answer, so the boundary has a hole per verb
- 2026-09-08 · RG65: resolving an engine costs two interpreter starts on Windows, where the same file is spelled two ways
- 2026-09-08 · RG64: the fast suite and the one that spawns a real engine are one command, so every run pays for both
- 2026-09-08 · RG61: the app ships a high-severity advisory it did not choose, reached through the design system
- 2026-09-08 · RG60: no test ever starts the app, so the isolation the window configures is checked by reading it
- 2026-09-08 · RG59: the renderer has no content policy, so a script tag that reaches the page runs beside the bridge
- 2026-09-08 · RG58: nothing checks style or dead code, so three packages drift apart file by file with the gate silent
- 2026-09-08 · RG57: a main-process change is invisible until the dev run is killed, because only the renderer is watched
- 2026-09-08 · RG56: an agent working this backlog has no instruction file, so every session rediscovers the rules by reading code
- 2026-09-08 · RG55: there is no CI, so every assertion this plan rests on runs only where somebody remembers to run it
- 2026-09-08 · RG54: contrast and keyboard reach are asserted by nobody, so an accessibility defect ships looking like a style
- 2026-09-08 · RG53: a marker is an emoji drawn by whatever font the platform picked, so one status looks like three
- 2026-09-08 · RG52: there is one ground, so somebody who works in dark reads this app in light
- 2026-09-08 · RG51: strings are written into the components, so a second locale means reopening every screen
- 2026-09-08 · RG48: the claim that a service is a transport swap is asserted by nothing, so it is an intention
- 2026-09-08 · RG47: settings have nowhere to live, so roots and preferences are gone when the window closes
- 2026-09-08 · RG46: there is no executable, so the app runs only where its source tree and toolchain already are
- 2026-09-08 · RG45: nothing watches the governed files, so a write by an agent or a terminal leaves the screen stale
- 2026-09-08 · RG43: the claude binary is assumed to exist, so a machine without it fails at the moment work starts
- 2026-09-08 · RG42: what a session wrote to the governed files is not watched, so the screen shows the backlog from before it ran
- 2026-09-08 · RG41: a line is handed over without being taken, so two sessions can be started on the same task
- 2026-09-08 · RG40: a session's output is a stream nothing renders, so what the agent is doing is invisible while it does it
- 2026-09-08 · RG38: nothing starts Claude Code, so a task and the session that does it are two windows and a copied prompt
- 2026-09-08 · RG36: an input has no budget, so a field is typed to whatever length and refused after the sentence exists
- 2026-09-07 · RG79: an unreadable answer reports not JSON and drops stderr, where the engine explained itself in full
- 2026-09-07 · RG69: a subcommand is a verb with a space in it, and the verb table has no way to spell one
- 2026-09-07 · RG44: the renderer holds node powers, so the half a service would serve to a browser can delete a file
- 2026-09-07 · RG39: there is no design system, so every screen decides its own type, spacing and colour on the spot
- 2026-09-07 · RG37: there is no application at all: no window, no build and no way to run any of this
- 2026-09-07 · RG35: amend, restate and renumber have no door, so the corrections that keep an id are terminal-only
- 2026-09-07 · RG34: the argv the app composed is never shown, so a write cannot be read before it runs or repeated after it
- 2026-09-07 · RG33: a lint finding is a line of text, so a report naming its own doors arrives as something to read and not to run
- 2026-09-07 · RG32: a rationale section cannot be written, so a line filed here points at nothing until a terminal is opened
- 2026-09-07 · RG31: there is no door for ship, retire, defer or resume, so a task is read here and closed elsewhere
- 2026-09-07 · RG30: a marker cannot be moved, so starting work is a shell command run beside the window that shows it
- 2026-09-07 · RG29: the app reads and never writes, so a correction is typed in the terminal this app exists to replace
- 2026-09-07 · RG27: the ledger and the decisions file are unreachable, so what shipped and what was weighed live outside this app
- 2026-09-07 · RG26: the non-goals and the criteria are off screen, so the two lists that bind a proposal are the two nobody reads
- 2026-09-07 · RG25: deps and what a ship unblocks are a list of ids, so following a chain means reading ids one at a time
- 2026-09-07 · RG24: the rationale section is not shown, so the half of a task that explains it stays invisible
- 2026-09-07 · RG23: a task has no detail, so the why, the deps and the design are read in the file this app exists to replace
- 2026-09-07 · RG22: a list of eight hundred lines has no filter, so a block or a marker is found by scrolling to it
- 2026-09-07 · RG21: opening a project shows counts and no lines, so the backlog it governs is still only in the file
- 2026-09-07 · RG20: finding a line by its words means opening each project and reading its list in turn
- 2026-09-07 · RG19: what to work on is asked one repository at a time, so there is no answer that ranges over all of them
- 2026-09-07 · RG18: whether a project's gate passes is unknown until it is opened, so the list cannot say which one needs attention
- 2026-09-07 · RG17: twenty projects are read one after another, so the first screen is a wait with nothing drawn on it
- 2026-09-07 · RG16: nothing shows more than one project, which is the single thing this app exists to do
- 2026-09-07 · RG14: the whole tree is walked again on every launch, so a list the person already approved is rediscovered
- 2026-09-07 · RG12: a git worktree family reads as unrelated projects, so one backlog appears once per version folder
- 2026-09-07 · RG11: a scan descends into node_modules and .git, so finding seventeen projects reads a hundred thousand folders
- 2026-09-07 · RG10: no root is named, so there is nowhere to look and the project list can only come back empty
- 2026-09-07 · RG8: a read has no ceiling, so one project whose engine hangs stops the screen that was drawing it
- 2026-09-07 · RG7: every read costs an interpreter start, so a screen over twenty projects pays twenty of them to redraw
- 2026-09-07 · RG6: nothing checks which build answered, so a project on an older roadkeep fails at the first flag it lacks
- 2026-09-07 · RG5: a refusal is prose on stderr, so the field it is about has to be read back out of English
- 2026-09-07 · RG4: the client's reading of a payload is asserted by nothing, so a rename upstream breaks a user and not a build
- 2026-09-07 · RG3: a payload arrives as any, so a key renamed upstream is an undefined at run time and never a red build
- 2026-09-07 · RG2: the engine is assumed to be a roadkeep on PATH, which is the one thing engines says it may not be
- 2026-09-07 · RG1: nothing here invokes roadkeep, so every payload this design rests on is one no code can fetch
- undated · RG278: a session card moves only under a pointer drag, and an arrangement nobody wants is undone one card at a time
- undated · RG62: nothing stops a screen re-declaring a component the design system already exports
- undated · RG49: the executable is unsigned, so Windows tells a person the app is untrusted before it opens